Research map / Vision-language and multimodal / Multimodal LLMs
Spatial Representation Distillation and Knowledge Routing for Vision-Language-Action Models
Jungin Park, Chaoran Zhu, Changjae Oh, Kwanghoon Sohn
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes spatial representation distillation and knowledge routing to inject 3D spatial knowledge into vision-language-action models without explicit 3D inputs.
Problem
VLAs need 3D spatial understanding for robot manipulation, but explicit 3D inputs (depth, point clouds) are costly and implicit spatial injection into visual features is limited.
Approach
Distills spatial representations into the VLA's visual features and uses a knowledge routing mechanism to direct spatial knowledge to the action pathway; based only on the abstract, the precise mechanism is not confirmed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add attention response distillation from VGGT to the LoRA vision branch (ablation) | SPARK-VLA (no SigLIP) · LIBERO average | success rate (%): 58.7 -> 94.1 | Improved |
| Add knowledge router (ablation) | SPARK-VLA · LIBERO average | success rate (%): 94.1 -> 98.5 | Improved |
| Use SigLIP together with DINOv2 (ablation) | VLA with vision LoRAs + projector · LIBERO average | success rate (%): 58.7 -> 72.4 | Improved |
| Train projector from scratch in addition to LoRAs (ablation) | LoRA-only VLA · LIBERO average | success rate (%): <5 -> 72.4 | Improved |
| Use attention response distillation instead of feature matching (ablation) | SPARK-VLA · LIBERO-Goal | success rate (%): 94.8 -> 98.0 | Improved |
| Use attention response distillation instead of attention-map matching (ablation) | SPARK-VLA · LIBERO-Goal | success rate (%): 96.4 -> 98.0 | Improved |
| Add geometric distillation instead of action loss only (ablation) | SPARK-VLA · LIBERO-Goal | success rate (%): 81.6 -> 98.0 | Improved |
| Use action-driven router instead of direct token concatenation (ablation) | SPARK-VLA · LIBERO-Goal | success rate (%): 94.0 -> 98.0 | Improved |
| Use router instead of FiLM fusion (ablation) | SPARK-VLA · LIBERO-Goal | success rate (%): 94.8 -> 98.0 | Improved |
| Use momentum moving average (0.9) for router targets (ablation) | SPARK-VLA router · LIBERO-Goal | success rate (%): 94.0 -> 98.0 | Improved |
| Increase momentum beyond 0.9 to 0.9999 (ablation) | SPARK-VLA router · LIBERO-Goal | success rate (%): 98.0 -> 97.0 | Hurt |
| Raise routing ratio rho from 0.7 to 1.0 (keep all tokens) (ablation) | SPARK-VLA router · LIBERO-Goal | success rate (%): 98.0 -> 94.0 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 4 on most shared results.
- VLANeXt — VLANeXt: Recipes for Building Strong VLA Models (ICML 2026) · better on 1 of 2 shared results
- VLA-Adapter — VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model (AAAI 2026) · better on 1 of 2 shared results
- Spatial Forcing — Spatial Forcing: Implicit Spatial Representation Alignment for Vision-language-action Model (ICLR 2026) · better on 0 of 1 shared result
- Evo-1 — Evo-1: Lightweight Vision-Language-Action Model with Preserved Semantic Alignment (CVPR 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Localizing, Structuring, and Rendering: Bridging 3D and 2D Vision-Language-Action Models for Robotic Manipulation (CVPR 2026 · similarity 0.91)
- ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation (CVPR 2026 · similarity 0.90)
- VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026 · similarity 0.90)
- From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.90)
- 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · similarity 0.90)
- GPT4Scene: Understand 3D Scenes from Videos with Vision-Language Models (ICLR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
